Re: [Qemu-devel] [PATCH V7 6/6] hostmem-file: add 'sync' option

[Yi Zhang <yi.z.zhang@linux.intel.com>]

On 2018-12-21 at 11:36:07 -0500, Michael S. Tsirkin wrote:
> On Fri, Dec 21, 2018 at 11:18:18AM +0800, Yi Zhang wrote:
> > On 2018-12-20 at 09:06:41 -0500, Michael S. Tsirkin wrote:
> > > On Thu, Dec 20, 2018 at 01:37:40PM +0800, Yi Zhang wrote:
> > > > On 2018-12-19 at 22:42:07 -0500, Michael S. Tsirkin wrote:
> > > > > On Thu, Dec 20, 2018 at 11:03:12AM +0800, Yi Zhang wrote:
> > > > > > On 2018-12-19 at 10:59:10 -0500, Michael S. Tsirkin wrote:
> > > > > > > On Wed, Dec 19, 2018 at 05:10:18PM +0800, Yi Zhang wrote:
> > > > > > > > > > +
> > > > > > > > > > + - 'sync' option of memory-backend-file is not 'off', and
> > > > > > > > > > +
> > > > > > > > > > + - 'share' option of memory-backend-file is 'on'.
> > > > > > > > > > +
> > > > > > > > > > + - 'pmem' option of memory-backend-file is 'on'
> > > > > > > > > > +
> > > > > > > > > 
> > > > > > > > > Wait isn't this what pmem was supposed to do?
> > > > > > > > > Doesn't it mean "persistent memory"?
> > > > > > > > pmem is a option for memory-backend-file, user should know the backend
> > > > > > > > is in host persistent memory, with this flags on, while there is a host crash
> > > > > > > > or a power failures.
> > > > > > > > 
> > > > > > > > [1] Qemu will take necessary operations to guarantee the persistence.
> > > > > > > > https://patchwork.ozlabs.org/cover/944749/ 
> > > > > > > > 
> > > > > > > > [2] Host kernel also take opretions to consistent filesystem metadata.
> > > > > > > > Add MAP_SYNC flags.
> > > > > > > 
> > > > > > > OK so I'm a user. Can you educate me please?  
> > > > > > We suppose an administrator should know it, what is the back-end region coming from,
> > > > > > is it persistent? what is the font-end device is? a volatile dimm or an
> > > > > > nonvolatile dimm? then they can choice put the pmem=on[off] and sync=on[off].
> > > > > > If he didn't, we encourage that don't set these 2 flags.
> > > > > > 
> > > > > > > When should MAP_SYNC not
> > > > > > > be set? Are there any disadvantages (e.g.  performance?)?
> > > > > > Not only the performance, sometimes like the front-end device is an
> > > > > > volatile ram, we don't wanna set such option although the backend is a
> > > > > > novolatile memory, if power lose, all of thing should lose in this ram.
> > > > > 
> > > > > 
> > > > > 
> > > > > I am not sure how does above answer the questions. If I don't know,
> > > > > neither will the hypothetical administrator. Looks like a better
> > > > > interface is needed to make the choice on behalf of the user.
> > > > > 
> > > > By default, we have already ignore the 2 flags, unless the administrator
> > > > know how to make that use. By-now, seems we don't have a better way to detect what
> > > > memory-backend-file is, a persistent memory or not.
> > > 
> > > In that case how about an interface where user tells QEMU "this backend
> > > is in persistent memory"?
> > 
> > Actually, [pmem=on] already did this, we can get the backend memory type from:
> > 
> > file_memory_backend_get_pmem(),
> > 
> > That is already being used in the memory_region_init_ram_from_file. that
> > is why I reuse the RAM_PMEM to identify the region coming from a
> > persitent memory? correct me if something I misunderstood? 
> 
> Right and thus my question: why not set SYNC unconditionally with pmem=on?
A case is that we set pmem but don't wanna sync in a /dev/dax backend,
which is direct mapping to physical nvdimm device without filesystem aware,
SYNC is useless in this condition.
> 
> > 
> > > 
> > > 
> > > 
> > > > > 
> > > > > > > 
> > > > > > > -- 
> > > > > > > MST
> > > > > > > 
> > > > > 
> > > 
>